ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
  
  
  
  
  
  
  
  
  
  
  
    
    
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
(a) eee a LA (a) (b) 
1 i 
i 1 1 i 
| Ll | 
i : 1 ' B 
i : ! i B 
| LB ! ! ' ; 
i : ! i W 
! I ; ! 
1 W B 
1 1 1 ^ 
| Eod : B B 
Ue cec m Um Ue ue Um m e e n 1 
! 1 
i i 
i i 
! i 
i 1 
Be ar te se 
CUN A+ res d) 
| | | 
i » ! i i 
1 Te; : | B B 
ra > HN lo snr. B n 
n ni T : | 
" TE à i i / 
I" i i "i ui W W 
141 ih uy " 
ET, th | "ni n! B 
Fa Et 1 n "i 
jn 8 | "i - B ] 
Hi ii E i ; 
a tho "i u 
ET. 1^ ] " i u 3 
n Ih o! nd An 
In In à n! 
141 a! 1 Wu i uo! 
ut ik j-r--1--------1" | 7 
Hl eed beet W 
J e—— tre nee timntion E i i 
i i : i 
i ' i i 
b--T------------ i i s ; 
kommen! Figure 5. Finding ICOs by a successive narrowing 
The neighbour set of Figure 2: h, i, j is used. 
Figure 4: Effect of expanding the neighbor set (a) Segmented image; (b) narrowing by using the neighbour of B; (c) 
(a), (b) and (c) Expanded neighbour set (dotted line); (d) the core of the narrowing by using the neighbour of G; (d) narrowing by using the 
expanded neighbour set fully enclose the fastened layout of O. subs; (e) neighbour of W. 
and (f) the expanded core can also contain unfastened layout of O. subs. 
Finding Objects: An object finding function can be designed (a) 
using Theorem 2. Suppose we seek an object O. Theorem 2 
states that we can determine O.area by logically conjuncting 
(b) 
every element of O.nset, ie, MBR (c,O.subs) for all 
ceO.subs. Therefore, we select a first candidate component 
c cO.subs in  SEG(I) and find the next component 
€) € O.subs in SEG(I) ^ MBR' (c,O.subs). 
  
The MBR' (cj, O.subs) used here is defined in MS. Repeating 
  
this process, we finally find O.area containing all elements of 
O.subs. This process is referred to as narrowing. Figure 5 
illustrates object finding via narrowing. : 
Area\Aspect R. Square Long Very Long 
Coping with Object Shape Variation: In real images, 
instances of O having the same O.attriv may have different 
Small Type A ype E Type C 
O.area.size due to intrinsic size variations and/or the scale Medium Type D Type E Type C 
change of the image O. This problem can be addressed using 
: T ; : 1 Large Type E "ype F Type G 
Theorem 3, which states that if a revised O.nser containing an d 
  
expanded MBR' (c,O.subs) is used, we can trap the O.area in . P ENS 
Figure 6. Classification of component shapes 
Qum : 
the core coups MBR (c, O.subs). (a) and (b) The minimum area MBR among vertical and the diagonal 
Therefore, expanded MBR (c,O.subs) should be used in (45^) candidates is useful to enhance the approximation accuracy of 
defining the model set MS. After a successive narrowing, we the component shape (area and aspect ratio); (c) classification of 
components by area and aspect ratio. This classification is effective for 
can determine the MBR of the found O.subs as an estimation OS ded à 
computerization of infinitely varying shapes. 
of O.area. Furthermore, it is often the case that objects in an 
image / are rotated from their models in MS. We can cope with Coping with Component Variations: Instances of O having 
this problem by extending the MS to MS” containing rotated the same O.attriv may not have identical O.subs due to 
models. component variations. Typical variations include changes in 
size and component deletions/additions. The size change 
problem is addressed by adding shape attributes to each 
A - 390